Nvidia books 2.5x substrates and 300% more interposer capacity

Nvidia is paying suppliers in advance to add 2.5 times more chip substrate capacity over three years and more than 300% more interposer capacity by the end of 2028, which shows that the parts after the wafer, not the wafer itself, are now the step Nvidia is spending its own cash to unblock.

What happened?

• The announcement: On 21 October 2026, at its GTC Berlin event, Nvidia announced capacity plans with six suppliers: IBIDEN, Samsung Electro-Mechanics, Unimicron, Amkor, King Yuan Electronics (KYEC) and Siliconware Precision Industries (SPIL). Chief executive Jensen Huang said: "Demand is extraordinary. Capacity is the constraint."
• Substrates: Nvidia plans to increase substrate capacity for its products by 2.5 times over the next three years. It is giving IBIDEN a multibillion-dollar capacity reservation advance for its Ono plant in Japan, which targets three times current capacity, with new capacity ramping from the first quarter of 2028. Samsung Electro-Mechanics joins Nvidia's substrate supply from existing plants in South Korea, with new factories in Sejong and Busan. Unimicron's expansion in Taiwan, backed by a capacity reservation agreement, starts adding capacity in the second quarter of 2027 and reaches full capacity by mid-2028.
• Packaging, test and optics: Amkor and Nvidia target a more than 300% increase in interposer wafer capacity in South Korea by the end of 2028. Nvidia expects its total chip testing capacity to double from 2026 to 2027, including a $1.1 billion capacity reservation advance to KYEC for new US test facilities (high-volume production targeted for April 2027) and expanded testing in Singapore (high-volume production expected from February 2027). SPIL targets 15 times more optical-engine capacity for co-packaged optics in Taiwan by mid-2027, and aims to double it again by the end of 2028.

Why does it matter?

The direct signal. A chip designer is putting its own money into the factories that come after the wafer. A capacity reservation advance is a prepayment: Nvidia pays now and gets a guaranteed share of the new output later. Nvidia does not say how large the IBIDEN advance is beyond "multibillion-dollar", and only the KYEC figure, $1.1 billion, is given in full.

The second-order signal: the back end of the chip is the stuck step, so the buyer is funding it. An AI GPU is not finished when TSMC makes the die. It still needs an interposer (the silicon sheet that joins the GPU to its memory stacks; see What Is an Interposer?), a substrate (the layered board the whole package sits on), testing, and, for the next generation of networking, optical engines packed beside the chip. Each of those is a separate factory run by a different company, and each can cap how many finished GPUs ship in a quarter. Nvidia's release names all four. That fits what we have tracked on TSMC's CoWoS line and in If Substrates Stay Short Through 2028, Where Does Cash Stick?: wafers have been ramping faster than the packaging around them.

Who carries the risk. Prepayments move part of the cost of a new plant from the supplier to Nvidia. The supplier still owns the building and the depreciation if demand falls, but it starts with a paying customer. Chip companies are doing the same further down the stack: on 13 October Fermi said AMD had signed a credit support agreement behind the 222 MW first phase of its Project Matador data center campus in Texas, to help Fermi raise project financing. The pattern is clear: the chip vendors with the most cash are now paying to clear the steps that would otherwise hold back their own sales.

Why substrates take years. Our arithmetic: 2.5 times in three years is about 36% more capacity each year, compounded. IBIDEN's extra Ono capacity starts ramping in early 2028 and Unimicron's reaches full capacity by mid-2028, so most of the substrate growth lands in the second half of the three years, not the first.

What does it change?

Near term (to mid-2027)

• Packaging stays tight. The first dated additions are Singapore testing in February 2027, Unimicron substrates from the second quarter of 2027 and KYEC's US testing in April 2027. Samsung Electro-Mechanics' existing capacity is the only one Nvidia describes as immediate. Until then, the number of finished GPUs still depends on the capacity that exists today.
• Testing becomes a named step. Doubling test capacity in a year (from 2026 to 2027) tells us that testing, not only packaging, has been holding back output. Test happens after packaging, so a GPU that cannot be tested cannot ship.
• Suppliers get visible demand. For Amkor, IBIDEN, Samsung Electro-Mechanics, Unimicron, KYEC and SPIL, a named Nvidia programme with prepaid capacity is easier to finance and staff than a forecast.

Longer term (2028 and after)

• More than four times the interposers. Our arithmetic: a more than 300% increase means Amkor's interposer wafer capacity in South Korea ends 2028 at more than four times today's level. That gives Nvidia a larger second source next to TSMC's own CoWoS lines; we compare the two routes in Who Makes Advanced Packaging?.
• Optics move into the package. Our arithmetic: 15 times by mid-2027 and another doubling by the end of 2028 is about 30 times today's optical-engine capacity at SPIL. That is a bet that co-packaged optics, where the light-to-electric converter sits next to the switch chip instead of in a plug-in module, moves from pilot to volume within two years.
• What we cannot see yet: the total size of the prepayments, today's capacity at each supplier (so the 2.5 times and 300% figures have no unit), how much of the new capacity is reserved only for Nvidia, and whether the materials inside substrates, such as specialist glass cloth, grow as fast.

What should we watch next?

• 26 October 2026: Amkor third-quarter results, after the US market close, with the call at 5:00 pm Eastern time (05:00 SGT on 27 October). Watch for the size of the South Korea interposer build, how much of it Nvidia prepays, and any change to Amkor's spending plan.
• 3 November 2026: Supermicro first-quarter results call, 5:00 pm Eastern time (06:00 SGT on 4 November). Supermicro builds the servers these packaged GPUs go into; watch whether it says chip or component supply limited its shipments.
• February and April 2027: the first dated test capacity, in Singapore and then at KYEC's new US site. Slips there would show up as GPU supply limits a quarter later.
• What would weaken the catalyst: a delay to the Unimicron or IBIDEN ramps, Nvidia cutting the prepayments, or a slowdown in AI server orders that leaves the new lines under-used.

FAQ

What did Nvidia announce on 21 October 2026?
Capacity plans with six suppliers: 2.5 times more substrate capacity over three years (IBIDEN, Samsung Electro-Mechanics, Unimicron), more than 300% more interposer wafer capacity at Amkor by the end of 2028, double the chip testing capacity from 2026 to 2027 (including KYEC in the US and new capacity in Singapore), and 15 times more optical-engine capacity at SPIL by mid-2027.

What is a capacity reservation advance?
A prepayment from a customer to a supplier. The supplier uses the cash to build or equip a plant, and the customer gets a reserved share of the output. Nvidia says it is giving one to IBIDEN (multibillion-dollar) and to KYEC ($1.1 billion), and has a capacity reservation agreement with Unimicron.

Does this end the AI chip packaging shortage?
Not soon. The earliest dated additions are in February and the second quarter of 2027, and the largest substrate additions ramp in 2028. Huang's own words were that capacity is the constraint.

Why does Singapore appear in the release?
Nvidia says expanded testing operations in Singapore are expected to reach high-volume production in February 2027, as part of doubling its total test capacity.

Takeaway: Nvidia is paying up front for substrates, interposers, testing and optics because those steps, not the wafer, now set how many AI chips ship, and most of that new capacity does not arrive until 2027 and 2028.

References

1. Nvidia, "NVIDIA and Supply Chain Partners Expand Capacity to Accelerate AI Infrastructure Buildout", 21 Oct 2026. All supplier, capacity, timing and prepayment figures, and the Jensen Huang quote.
2. Amkor Technology, "Amkor Technology to Announce Third Quarter 2026 Financial Results on October 26, 2026", 5 Oct 2026. The results date and call time.
3. Supermicro, "Supermicro Provides First Quarter of Fiscal Year 2027 Preliminary Business Update", 20 Oct 2026. The revenue estimate, prior guidance and 3 November call.
4. Fermi, "Fermi Signs Credit Support Agreement with AMD and Announces Partnership", 13 Oct 2026. AMD's credit support for the 222 MW first phase of Project Matador.
5. ETElectronicsWorld, "TSMC plans 3nm chip manufacturing in second Japan factory", 21 Oct 2026. TSMC's statement on the second JASM fab.
6. Newspim, "Hyein to supply 46.3 billion won of standby generators to Pohang AI data center" (in Korean), 21 Oct 2026. The 20 generators, 61.5 MW, contract value and May 2027 deliveries.

Not investment advice. The yearly substrate growth rate, the four-times interposer figure and the 30-times optical-engine figure are our arithmetic from the sources above.